CSAIL Event Calendar

Finding Diverse M Best Solutions in a Random Field

Much effort has been directed at algorithms for obtaining the highest
probability (MAP) configuration in a probabilistic (random field)
model. In many situations, one could benefit from additional solutions.
Unfortunately, current methods for computing additional most
probable configurations (M-Best MAP) produce solutions that tend to
be very similar to the MAP solution and each other, thus lacking diversity.
In this talk I will describe an algorithm for finding a diverse set of
highly probable solutions under a discrete probabilistic model. Given a
dissimilarity function measuring difference between two solutions, our
algorithm minimizes a linear combination of the energy under the model
and similarity to previously found solutions. We show that for certain
widely applicable families of dissimilarity functions we can guarantee
that these solutions can be found as easily as the MAP solution. I will
describe how this approach can be used to significantly improve
performance in three applications in computer vision: interactive
image segmentation, multi-category image labeling, and human pose
tracking in video.

Joint work with Payman Yadollahpour, Dhruv Batra and Abner Guzman.

See other events happening in July 2012